Optimization

Objective

The basic objective of any optimization method is to find the values of the system
state variables and/or parameters that minimize some cost function of the system.

Some examples:

Minimize:
• the error between a set of measured and calculated data,
• active power losses,
• the weight of a set of components that comprise the system,
• particulate output (emissions),
• system energy, or
• the distance between actual and desired operating points
Formulation

Minimize f ( x, u ) x∈R n

u ∈ Rm

Subject to
g ( x, u ) = 0 (equality constraints) Power flow equations

h( x, u ) = 0 (inequality constraints) Transmission line limits, Qgen limits, etc

where x is the vector of system states and u is the vector of system parameters.

PQ bus voltage magnitudes and phase PV bus voltage magnitudes, Pgen

Least Squares Estimation (LSE)

State Estimation is the process of estimating unknown states from measured

quantities.

In many systems, more measurements are made than are necessary to uniquely
determine the operating point.

Conversely, not all of the states may be available for measurement.

State estimation gives the “best estimate” of the state of the system in spite of
uncertain, redundant, and/or conflicting measurements.
Let z be the set of measured quantities

The measured values of z may not equal the true values of ztrue due to
• measurement errors,
• missing data, or
• intentional misdirection (hacking)

We’d like to
• Identify bad data
• Minimize the error between the true and measured values

e =−
z z true =−
z Ax
States of the system
 “squared error”

2
 m m
Minimize e = e ⋅ e = ∑  zi − ∑ aij x j 
2 T

=i 1 =  j 1 

Minimizing the square of the error is desired to negate any effects of sign
differences between the measured and true values.

( z − Ax ) ( z − Ax )
T
e ⋅e =
T

Minimize
(
= z T
− x A ) ( z − Ax )
T T

=z z − z Ax − x A z + x A Ax
T T T T T T

=
z T z − 2 xT AT z + xT AT Ax
Minimize z T z − 2 xT AT z + xT AT Ax

Take derivate and set equal to zero:

∂ T
∂x
(
z z − 2 xT AT z + xT AT Ax = )
−2 AT z + 2 AT Ax =
0

Solving for x:
AT Ax = AT z Solve using LU factorization (or other linear solver)

ˆ = bˆ
Ax

( )
−1
x= A A T
AT z Analytic solution
 R=
1 R=
3 R=
5 1.5Ω
Find the node voltages R=
2 R=
4 1.0Ω Circuit equations
V1 and V2 −V1 + R1 z1 + z3 =0
−V2 − R5 z2 + z4 =
0
z3 / R2 − z1 + ( z3 − z4 ) / R3 =
0
z4 / R4 + z2 + ( z4 − z3 ) / R3 =
0
Yielding:
 z1   0.4593 −0.0593
 z   0.0593 −0.4593 V
 2 =   1
 z3   0.3111 0.0889  V2 
   
 4 
z 0.0889 0.3111

( )
−1
Least squares estimation: x = A A
T
AT z Analytic solution

−1
  0.4593 −0.0593 T
 0.4593 −0.0593   0.4593 −0.0593
T
 z1 
   0.0593 −0.4593   0.0593 −0.4593 z 
V1    0.0593 −0.4593      2
V  =   0.3111 0.0889   0.3111 0.0889    0.3111 0.0889   z3 
 2 
 0.0889        
0.3111 0.0889 0.3111 0.0889 0.3111  z4 
 
Solving for voltages: V1 = 9.8929
V2 = 5.0446

Finding the error: e= z − Ax

 4.27   0.4593 −0.0593
 −1.71  0.0593 −0.4593 9.8929
     
= −
 3.47   0.3111 0.0889  5.0446 
   
 2.50   0.0889 0.3111 
 0.0255
 0.0205
=  How do we know if this is a reasonable error?
 −0.0562 
 
 −0.0512 
Weighted Least Squares Estimation (WLSE)

Suppose we have more confidence in certain types of measurements:

• Certain sensors have been more recently calibrated
• Some types of sensors tend to be more accurate (smaller standard deviation)

We’d like to apply more weight to

measurements in which there is
greater confidence
Introduce a weighting factor:
2
m  m 
minimize e = e ⋅ e = ∑ wi  zi − ∑ aij x j 
2 T

=i 1 =  j 1 

In matrix form:

( A WA)
−1
T
= A Wz
A WAx T
⇒ =
x T
AT Wz
Let 100 0 0 0
 0 100 0 0 
W =  The currents are weighted more
 0 0 50 0  heavily than the voltages
 
 0 0 0 50 

 0.0141 Decreased error in current

 0.0108
V1  9.9153
Resulting in   =   with e =  
 2 
V 5.0263   −0.0616 
 
 0.0549   0.0255
 0.0205
V1 = 9.8929  
original  −0.0562 
V2 = 5.0446
 
 −0.0512 
A plausible weighting matrix that reflects the level of confidence in each
measurement set is the inverse of the covariance matrix W = R−1.

 1 
σ 2 0 … 0 
 1 
 1 
 0 … 0 
= −1
W R=  σ 22 
     
 
 0 1 
0 …
 σ m2 

Measurements that come from instruments with good consistency (small

variance) will carry greater weight than measurements that come from less
accurate instruments (high variance).
Bad Data Detection

Bad data may cause the accuracy of the estimate

to deteriorate
To determine whether the estimated values
differ significantly from the measured values, a
useful statistical measure is the χ2 (chi-squared)
test of inequality.
By comparing the weighted sum of errors with the
χ2 value for a particular degree of freedom and significance level, it can be
determined whether it is likely that bad data exists

A significance level of 0.05 indicates there is a 5% likelihood that bad data exist, or
conversely, a 95% level of confidence in the goodness of the data.
Test for Bad Data

1. Use z to estimate x

2. Calculate the error

e = z − Ax
3. Evaluate the weighted sum of squares
m 2
1
f =∑ ei2
i =1 σi
4. For k = m − n and a specified probability α, if f < χ k2,α then the
data are good; otherwise at least one bad data point exists.
 Using the chi-square test of inequality with α=0.01, check
for the presence of bad data

The number of states is 2 and the number of measurements

is 4: k = 4-2 = 2.

m=4
1
f =∑ ei2
i =1 σi
= 100(0.0141) 2 + 100(0.0108) 2 + 50(−0.0616) 2 + 50(0.0549)
= 0.3720

< 9.21 Data good (with 99% confidence)

Nonlinear Least Squares Estimation

∑ w  z − h ( x )
2
Minimize e 2
= e ⋅e =
T
i i i
i =1

Nonlinear equations

The approach to finding the states (and inputs) that minimize the weighted,
squared error is similar to the linear case.
The state values that minimize the error are found by setting the derivatives of
the error function to zero:

) H xT R −1 [ z − h( x=
F ( x= )] 0
where solve using a nonlinear solution
 ∂h1 ∂h1 ∂h1  method such as Newton-
 ∂x …
∂xn 
Raphson
∂x2
 1 
 ∂h2 ∂h2 ∂h2 
 …
H x =  ∂x1 ∂x2 ∂xn 

     
 
 ∂hm ∂hm
…
∂hm 
 ∂x1 ∂x2 ∂xn 
Using Newton-Raphson to solve the nonlinear system of equations requires the
Jacobian matrix:
∂
J F ( x) = T
H ( x) R [
−1
z − h ( x ) ] =
− H T
( x ) R −1
H x ( x)
∂x
x x

Combining with the NR iteration yields the iteration:

not just this

 ( ) ( )
 H xT x k R −1 H x x k   x=

k −1
− x k

 H T
x x ( )
k
R −1

 z − h ( x k
) 

whole expression must be ≈ 0 for convergence
 Linear Programming

• Linear programming is one of the most common forms of optimization.

• Linear programming can be used when a problem can be expressed by a linear
objective (or cost) function to be maximized (or minimized) subject to linear
equality or inequality constraints.

minimize f ( x) = cT x
subject to Ax ≤ b
x≥0
• Any linear programming problem described by (A, b, c), there exists
another equivalent, or dual problem (−AT ,−c,−b).
 feasible region

minimum cost

The simplex method is an organized search of the vertices by moving along the
steepest edge of the polytope
Comments:

1. maximizing c x is the same as minimizing −c x

T T

2. any constraint of the form a x ≥ β is equivalent to − a x ≤ − β

T T

3. any constraint of the form a x = β is equivalent to a x ≤ β and

T T

− aT x ≤ − β
4. if a problem does not require xi to be nonnegative, then xi can be replaced by
the difference of two variables xi = ui ‒ vi where ui and vi are nonnegative.

If a linear programming problem and its dual both have feasible points (i.e.
any point that satisfies Ax ≤ b, x ≥ 0 or −AT y ≤ −c, y ≥ 0 for the dual
problem), then both problems have solutions and their values are the
negatives of each other.
Simplex Method

The simplex method is often accomplished by representing the problem in tableau

form, which is then modified in successive steps according to given rules.

The tableau has the following initial form:

cT 0 0

A I b

0 bT
Rules:

• The x vector must satisfy the equality constraints Ax = b

• The x vector must satisfy the inequality x ≥ 0
• There are n components of x (designated nonbasic variables) that are
zero. The remaining m components are usually nonzero and are
designated as basic variables.
• In the matrix that defines the constraints, each basic variable occurs in
only one row.
• The objective function f(x) must be expressed only in terms of nonbasic
variables.
• An artificial variable may be added to one or more constraints to obtain a
starting solution.
Simplex Algorithm

1. If all coefficients in f(x)=cTx (i.e. top row of the tableau) are greater than or
equal to zero, then the current x vector is the solution.

2. Select the nonbasic variable whose coefficient in f(x) is the largest negative
entry. This variable becomes the new basic variable xi.

3. Divide each bi by the coefficient of the new basic variable in that row, aij. The
value assigned to the new basic variable is the least of these ratios
x j = bk / akj
4. Using pivot element akj, create zeros in column j of A with Gaussian
elimination. Return to 1.
Example:

Minimize:
f ( x) : − 6 x1 − 14 x2
Subject to:
2 x1 + x2 ≤ 12
2 x1 + 3 x2 ≤ 15
x1 + 7 x2 ≤ 21
x1 ≥ 0, x2 ≥ 0
feasible region
Rewrite:

Minimize:
f ( x) : − 6 x1 − 14 x2 − 0 x3 − 0 x4 − 0 x5
Subject to:
2 x1 + x2 + x3 =
12 basic variables

2 x1 + 3 x2 + x4 =
15
x1 + 7 x2 + x5 =
21
x1 ≥ 0, x2 ≥ 0, x3 ≥ 0, x4 ≥ 0, x5 ≥ 0
 Tableau c Tx
I

-6 -14 0 0 0 0 b
2 1 1 0 0 12
A 2 3 0 1 0 15
1 7 0 0 1 21
0 0 12 15 21

x1 = 0
x2 = 0
x3 = 12
x4 = 15 f ( x) =
−6 x1 − 14 x2 =
0
x5 = 21
 x1 x2 x3 x4 x5 f(x)
-6 -14 0 0 0 0
2 1 1 0 0 12
2 3 0 1 0 15
1 7 0 0 1 21
0 30 9
12 6
15 0
21

Increasing
Set x2 x=2 3will decrease
(with andmost
x1 = 0)f(x) rapidly, therefore hold x1 constant at zero and let
solve:
x2 increase.
x3 = 12 − x2 = 9
2 x1 + x2 + x3 = 12 0 ≤ x3 = 12 − x2 x2 ≤ 12
x4 =15 − 3 x2 =6
2 x1 + 3 x2 + x4 =15 0 ≤ x4 = 15 − 3 x2 x ≤ 5
x5 =21 − 7 x2 =0 x5 is now a non-basic variable and x2 2
is a basic variable
x1 + 7 x2 + x5 = 21 0 ≤ x5 = 21 − 7 x2 x2 ≤ 3
most restrictive
 x1 x2 x3 x4 x5 f(x)
-4
-6 -14
0 0 0 20 -42
0
2 1 1 0 0 12
2 3 0 1 0 15
1 7 0 0 1 21
0 3 9 6 0

Must express f(x) in terms of non-basic variables:

(21 − x5 − x1 )
x2 = f ( x) =
−6 x1 − 14 x2
7
(21 − x5 − x1 )
=
−6 x1 − 14
7
=
−4 x1 + 2 x5 − 42 ⇒ f ( x) =
−42
 x1 x2 x3 x4 x5 f(x)
-4 0 0 0 2 -42
2 1 1 0 0 12
2 3 0 1 0 15
1 7 0 0 1 21
0 3 9 6 0

Need to zero out below x2 with Gaussian elimination:

(21 − x5 − x1 ) 13 1
2 x1 + + x3 ≤ 12 x1 − x5 ≤ 9
7 7 7
(21 − x5 − x1 ) 11 3
2 x1 + 3 + x4 ≤ 15 x1 − x5 ≤ 6
7 7 7
doesn’t change
x1 + 7 x2 + x5 ≤ 21 1 1
except to normalize x1 + x2 + x5 ≤ 3
7 7
 x1 x2 x3 x4 x5 f(x)
-4 0 0 0 2 -42
13/7 0 1 0 -1/7 9
11/7 0 0 1 -3/7 6
1/7 1 0 0 1/7 3
0 3 9 6 0

Need to zero out below x2 with Gaussian elimination:

(21 − x5 − x1 ) 13 1
2 x1 + + x3 ≤ 12 x1 − x5 ≤ 9
7 7 7
(21 − x5 − x1 ) 11 3
2 x1 + 3 + x4 ≤ 15 x1 − x5 ≤ 6
7 7 7
x1 + 7 x2 + x5 ≤ 21 1 1
x1 + x2 + x5 ≤ 3
7 7
 x1 x2 x3 x4 x5 f(x)
-4 0 0 0 2 -42
13/7 0 1 0 -1/7 9
11/7 0 0 1 -3/7 6
1/7 1 0 0 1/7 3
0 3 9 6 0

Return to Step 1.
The cost function will only decrease for an increase in x1: set x5 to zero and let x1 increase.

13 63
New constraints: 0 ≤ x3 =9 − x1 x1 ≤
7 13
11 42
0 ≤ x4 =6 − x1 x1 ≤ most restrictive
7 11
0 ≤ 7 x2 =21 − x1 x1 ≤ 21
 x1 x2 x3 x4 x5 f(x)
-4 0 0 0 2 -42
13/7 0 1 0 -1/7 9
11/7 0 0 1 -3/7 6
1/7 1 0 0 1/7 3
0 3 9 6 0

42 27
Set x1 = and solve for remaining variables: x2 =
11 11
21
x3 =
11
x=
4 x=
5 0 New nonbasic variables
 x1 x2 x3 x4 x5 f(x)
0
-4 0 0 28/11
0 2 -630/11
-42
0
13/7 0 1 -13/11
0 -1/7 21/11
9
1
11/7 0 0 7/11
1 -3/11
-3/7 42/11
6
0
1/7 1 0 -1/11
0 1/7 27/11
3
42/11
0 27/11
3 21/11
9 06 0

630
Rewrite f (x) in terms of nonbasic variables: f min ( x) = −
11
f ( x) =
−4 x1 + 2 x5 − 42 Both coefficients > 0, therefore no
more reduction in f (x) is possible
Solve for x1 in terms of x4:
f ( x) =
−4 x1 + 2 x5 − 42
7
=
x1 ( 6 − x4 ) =
28
x4 + 2 x5 −
630
11 11 11
Interior Point Method

feasible region

Iterations stay in the interior of the feasible region

Interior point methods construct a sequence of strictly feasible points that converges to
the solution stepsize
direction

x k +=
1
xk + α pk

One of the key aspects of the interior point method is that a transformation is applied
such that the current feasible point is moved (through the transformation) to the center
of the interior.

The new direction is then computed and the interior point is transformed back to the
original space. Each pk must be orthogonal to the rows of A
The projection matrix P
P= I − AT ( AAT ) −1 A
will transform any vector v into Pv=p

The transformation will also center the new iterate in the feasible space through scaling.

Scale x = Dxˆ so that 1 where D is the diagonal matrix.

1

xˆ = 1
k


 
1
Primal Affine Interior Point Method
1. Let k = 0.
2. Let D = diag(xk)
=
3. Compute Aˆ =
AD , cˆ Dc
4. Compute Pˆ= I − A ˆ ˆ T ) −1 Aˆ
ˆ T ( AA
5. Set p = Pc
ˆˆ
k

6. Set θ = − min j p kj
α k
k +1
7. Compute xˆ = e + p
θ
k +1 k +1
8. Compute x = Dx
ˆ
9. If x
k +1
− x k ≤ ε then done. Else set k=k+1. Go to step 2.
Example:

Minimize:
f ( x) : − 6 x1 − 14 x2
Subject to:
2 x1 + x2 ≤ 12
2 x1 + 3 x2 ≤ 15
x1 + 7 x2 ≤ 21
x1 ≥ 0, x2 ≥ 0

(Same example system as before)

Rewrite with slack variables (as before):

Minimize:
f ( x) : − 6 x1 − 14 x2 − 0 x3 − 0 x4 − 0 x5
Subject to:
2 x1 + x2 + x3 =
12
2 x1 + 3 x2 + x4 =
15
x1 + 7 x2 + x5 =
21
x1 ≥ 0, x2 ≥ 0, x3 ≥ 0, x4 ≥ 0, x5 ≥ 0
 [ ]
Choose a feasible initial starting point: x o = 1 1 9 10 13 z 0 = cT x 0 = −20

The first scaling matrix is

1 
 1 
 
D= 9 
 
 10 
 13
The rescaled matrix  and objective function vector ĉ are computed as:
1 
2 1 1 0 0  1 
2 1 9 0 0
 
= =  2
Aˆ AD 3 0 1 0   9 =   2 3 0 10 0 
 
 
1 
7 0 0 1  10   1 7 0 0 13
 13
1   −6   −6 
 1   −14   −14 
    
= = 
cˆ Dc 9   0=   0
    
 10   0  0
 13  0   0 
The projection matrix P̂ is:
ˆ ˆ T ) −1 Aˆ
Pˆ= I − Aˆ T ( AA
−1
1   2 2 1   2 2 1 
 1   1 3 7  2 1 9 0 0  1 3 7 
       2 1 9 0 0
 1  −  9 0 0    2 3 0 10 0   9 0 0    2 3 0 10 0 
        
 1   0 10 0    1 7 0 0 13  0 10 0    1 7 0 0 13
 1  0 0 13   0 0 13 

 0.9226 −0.0836 −0.1957 −0.1595 −0.0260 
 −0.0836 0.7258 −0.0621 −0.2010 −0.3844 

=  −0.1957 −0.0621 0.0504 0.0578 0.0485
 
 −0.1595 −0.2010 0.0578 0.0922 0.1205
 −0.0260 −0.3844 0.0485 0.1205 0.2090 
The projected gradient is

p 0 = − Pc
ˆˆ
 0.9226 −0.0836 −0.1957 −0.1595 −0.0260   −6 
 −0.0836 0.7258 −0.0621 −0.2010 −0.3844   −14 

= −  −0.1957 −0.0621 0.0504 0.0578 0.0485  0 
  
 −0.1595 −0.2010 0.0578 0.0922 0.1205  0 
 −0.0260 −0.3844 0.0485 0.1205 0.2090   0 
 4.3657 
 9.6600 
 
=  −2.0435
 
 −3.7711 
 −5.5373
Calculate θ = − min j p 0j = 5.5373.
−1 0
=
Rescale the current iterate to x̂ 0 D= x e
and move to x̂ in the transformed space with α = 0.9:
1

1 1.7096 
1  2.5701
  
xˆ1 =1 + α p 0 =0.6679 
  
1  0.3871
1 0.1000 
Transforming this point back to the original space:

1.7096 
 2.5701
 
= =ˆ1  6.0108
x1 Dx
 
3.8707 
1.3000 

which has the cost function: cT x1 = −46.2383.

Continue until convergence.

 Same result as Simplex method

(1)
(2)

(0)
Consider the case in which there are no inputs and neglecting inequalities:

minimize f ( x) x ∈  n
subject to c( x) = 0

which becomes: minimize C * ( x) : f ( x) − λ c( x)

This is solved by taking the derivatives and setting them equal to zero:
∂C * ( x) ∂f ( x) ∂c( x)
=0 = −λ
∂x ∂x ∂x
∂C * ( x)
0= = −c( x) enforces equality constraint
∂λ
Example: Minimize C :
1 2
2
(
x + y2 )
Subject to: 2 x − y =5

The point that minimizes the constrained cost function is the

point in the x-y plane where the constraint line intersects the
circle defined by C. minimum

The minimum corresponds to the minimum radius of the

circle centered at the origin.
The augmented (Lagrangian) cost function is:

*
C :
1 2
2
( 2
)
x + y − λ ( 2 x − y − 5)

The solution is obtained by solving:

∂C * ∂C *
0= = x − 2λ 0= = 2x − y − 5
∂x ∂λ
∂C *
0= = y+λ
∂y
x=2
Resulting in y = −1 with C :
1
2
(
(2) + (−1) =
2 2 5
2
)
λ =1
If the cost function or equality constraints contain the input u, then an iterative
approach is used (neglecting inequalities)

minimize C ( x, u ) : f ( x, u ) − λ c ( x, u ) x ∈  u ∈ 
* T n m

a vector with dimension m

 ∂C * 
Taking derivatives:
 = 0= c( x, u )
 ∂λ 
 ∂C   ∂f   ∂c 
* T

 = 0=   −   [ λ ]
 ∂x   ∂x   ∂x 
 ∂C *   ∂f   ∂c 
T

 = 0=   −   [ λ ]
 ∂u   ∂u   ∂u 
Power System Applications
Optimal Power Flow

The optimal power flow problem is to formulate the power flow problem to find
system voltages and generated powers within the framework of the objective
function.

In this application, the inputs to the power flow are systematically adjusted to
maximize (or minimize) a scalar function of the power flow state variables.

The two most common objective functions are minimization of generating costs and
minimization of active power losses.
The case in which active power losses in the system are neglected, the optimal
power flow problem is typically called the Economic Dispatch problem.

The individual generating unit costs are given as quadratic functions of generating
power Pi

Minimize ∑ Ci ( Pi )=
i∈PV
∑ i 0 i1 i i 2 i
k
i∈PV
+ k P + k P 2

Subject to: ∑ P −P
i∈PV
i T =
0

Total system load

Introduce Lagrangian multipliers:
 
Minimize: ∑(
i∈PV
)
ki 0 + ki1 Pi + ki 2 Pi − λ  ∑ Pi − PT 
2

 i∈PV 
Taking derivatives:

0 =k11 + k12 P1 − λ
 k12 −1  P1   −k11 
0=k21 + k22 P2 − λ      
Linear system  k22 −1  P2   −k21 
 of equations       =   
0= km1 + km 2 Pm − λ     
 km 2 −1  Pm   −km1 
=
0 PT − ∑P
i∈PV
i  1 1  1 0   λ   PT 
Equal incremental cost rule

Note that

λ k11 + k12 P1
= λ is called the “incremental cost” and this problem solution
λ k21 + k22 P2
= is known as the “equal incremental cost” rule.


λ km1 + km 2 Pm
= The optimal solution occurs when the incremental cost for
each generator is the same. That implies that an
incremental change in any generator’s output will result in
the same incremental cost increase.
Example: Three generators with the following cost functions serve a load of 952 MW.
Assuming a lossless system, calculate the optimal generation scheduling.

C1 : P1 + 0.0625 P12 $/hr

C2 : P2 + 0.0125 P22 $/hr
C3 : P3 + 0.0250 P32 $/hr

0.125 0 0 −1  P1   −1 
 0 0.025 0 −1  P   −1 
Using the equal   2 =  
incremental cost rule:
 0 0 0.050 −1  P3   −1 
    
 1 1 1 0   λ  952 
Solving: P1 = 112 MW
P2 = 560 MW
P3 = 280 MW which yields a generating cost of $7,616/hr

λ = $15/MWhr

The value of λ is the incremental or break-even cost of production.

λ=
1 + 0.125 P1 =
1 + 0.025 P2 =
1 + 0.050 P3

The value of λ can provide insight into whether or not a utility should buy or
sell generation.
If a buyer is willing to pay $16/MW hr for generation, how much excess generation
should be produced and sold, and what is the profit for this transaction?

Rewriting: λ=
1 + 0.125 P1 =
1 + 0.025 P2 =
1 + 0.050 P3
P1 8 ( λ − 1)
=
P2 40 ( λ − 1)
=
P3 20 ( λ − 1)
=

Substituting λ= $16/MW hr into the generation equations: P1 = 120MW

P2 = 600MW
P3 = 300MW
The excess generation is:

P1 + P2 + P3 − 952 = 120 + 600 + 300 − 952 = 68 MW excess

The new cost of generation is:

C : P1 + 0.0625 P12 + P2 + 0.0125 P22 + P3 + 0.0250 P32 = $8670/hr

Amount received for sale of excess generation:

68 MW × $16 /MWhr =
$1088/hr

Therefore, the total cost is $8,670–1,088 = $7,580/hr. This amount is $34/hr less
than the original cost of $7,616/hr; thus, $34/hr is the profit achieved from the sale
of the excess generation at $16/MW hr.
Utility incremental cost table:

What incremental
corresponds to
10,000 MW? Total
Units generation
1222 1222
+160+240 1622
+240+310 2172
+320+380+220 3092
+390+450+410 4342
+470+520+590 5922
+540+587+608+100+20 7777
+587+110+130+30 8634
+170+160+50 9014
+230+200+60 9504
+502+290+230+80 10606
 From equal incremental cost rule

Using Pgen = [0.112 0.560 0.280]′ as a

starting point, find the optimal scheduling of
this system considering losses.

The cost functions of the generators are

C1 : P1 + 0.0625 P12 $/hr
C2 : P2 + 0.0125 P22 $/hr
C3 : P3 + 0.0250 P32 $/hr

Use the steepest descent method

Set up equations:
−1
 ∂f   ∂g 
T
  ∂g  T
  ∂f   P2  θ 2 
Gradient: ∇=
C  −       u=  x= 
 ∂u   ∂u    ∂x    ∂x   P3  θ3 
Why not P1?

Cost function: f : C1 + C2 + C3 =
P1 + 0.0625 P12 + P2 + 0.0125 P22 + P3 + 0.0250 P32

3
Equality constraints: g1 : 0 = P2 − PL 2 − V2 ∑ ViY2i cos (θ 2 − θi − φ2i )
i =1
3
g 2 : 0 = P3 − PL 3 − V3 ∑ ViY3i cos (θ3 − θi − φ3i )
i =1
Find matrices:

 ∂g  1 0 
 ∂u  = 0 1  ∂g1
    = V2 (V1Y12 sin (θ 2 − θ1 − φ21 ) + V3Y13 sin (θ 2 − θ3 − φ23 ) )
∂δ 2
 ∂g1 ∂g1 
∂g1
 ∂g   2
 ∂θ ∂θ3  =−V2V3Y32 sin (θ 2 − θ3 − φ23 )
 ∂δ 3
 ∂x  =  ∂g ∂g 2  System Jacobian
2 ∂g 2
 ∂θ ∂θ3  =−V3V2Y23 sin (θ3 − θ 2 − φ32 )
 2 ∂δ 2
∂g 2
= V3 (V1Y13 sin (θ3 − θ1 − φ31 ) + V2Y23 sin (θ3 − θ 2 − φ32 ) )
∂δ 3
Continuing:

 ∂f  1 + 0.025 P2 
 ∂u  = 1 + 0.050 P 
   3

 ∂f   ∂f   ∂P1   ∂f 
 ∂x  =  ∂P   ∂x    = 1 + 0.125 P1
 1   ∂P1 

P1 V1 (V1Y11 cos (θ1 − θ1 − φ11 ) + V2Y12 cos (θ1 − θ 2 − φ12 ) + V3Y13 cos (θ1 − θ3 − φ13 ) )

 ∂P1  V1V2Y12 sin (θ1 − θ 2 − φ12 ) 

 ∂x  = V V Y sin (θ − θ − φ ) 
   1 3 13 1 3 13 
 Per unit
Start deepest descent iteration
 P2  0.560 
=
Starting with u =    find the corresponding x values (from powerflow):
 P3  0.280 
θ 2  0.0279 
 ∂g  1 0  = x =   0.0128
=
 ∂u  0 1  θ
 3  
 
Calculate matrices:
 ∂g   −13.3048 9.9148
 ∂x  =  9.8849 −19.9960 
 
 ∂f  1.0140 
 ∂u  = 1.0140 
 
 ∂f   −3.3773   −3.4251 
= =
 ∂x  1.0141  −10.0853  −10.2278
   
 0.0048
Gradient: ∇C = 0.0021
 
With γ = 1  P2  0.560  0.0048 0.5552 
P  = 0.280  − γ  0.0021 =
0.2779 
 3      

Proceeding until convergences yields the final

powers:
 P1  136.6 
 P  = 517.9  MW
 2  
 P3   298.4  Higher than the
lossless case

with cost: $7,698/MWhr

Limitations on Independent Variables (inputs):

Let 80 ≤ P1 ≤ 1200 MW
450 ≤ P2 ≤ 750 MW
150 ≤ P3 ≤ 250 MW exceeded

If any Pi exceeds its limit, set it to its limit (i.e. P3 = 0.25 pu) and reduce the number of
inputs by 1
The new partial derivatives become:  P1  147 
 P  = 556  MW
 ∂g  1   ∂f 
 ∂u = [1 + 0.025 P2 ]
 2  
 ∂u  = 0   P3   250 
  solving

 ∂g   ∂f  with cost: $7,728/MWhr

=
 ∂x  = same  ∂x  same
Higher than the unconstrained case
Limitations on Dependent Variables (states):

Let P23 ≤ 100 MW (the powerflow in line 2-3 – originally 121.1 MW)

Two approaches:
1. Soft constraint – violating the limiting factor is undesirable
2. Hard constraint – violating the limiting factor is catastrophic
 Soft constraint: Use Penalty Functions

A penalty function is a function that is small for values less than the limit, but
increases rapidly if the limit is violated

=
p (h) e kh k > 0
=p (h) x 2 n e kh n, k > 0
p ( h) =
ax 2 n e kh + be kh n, k , a, b > 0

The new cost function becomes:

(
C ∗ : C (u , x) + λ T g (u , x) + p h(u , x) − h max )
Use a penalty function to enforce the constraint P23 ≤ 100 MW

Let p (h) = ( h − hmax ) where h = P23 = V2V3Y23 cos (θ 2 − θ3 − φ23 ) − V22Y23 cos φ23
2

The new partial derivatives are all the same except:

From penalty function
 ∂f   ∂f   ∂P1   ∂f   ∂P23 
= ∂x   ∂P   ∂x  +  ∂P   ∂x 
 1     23   
V1V2Y12 sin (θ1 − θ 2 − φ1,2 ) 
= (1 + 0.125 P1 )  
V1V3Y13 sin (θ1 − θ3 − φ1,3 ) 
 −V2V3Y23 sin (θ 2 − θ3 − φ23 ) 
+2 ( P23 − 0.1)  
 V2V3Y23 sin (θ 2 − θ3 − φ23 ) 
Proceeding with the steepest gradient algorithm iterations yields the final
constrained optimal generation scheduling:

 P1  141.3 
 P  =  491.9  MW with cost: $7,781/MWhr
 2  
 P3  319.7 

Higher than the unconstrained case

and P23 = 101.2 MW.

 Hard constraint: Dual Variable Approach

In this approach, the inequality constraints are added as additional equality

constraints with the inequality set equal to the limit (upper or lower) that is
violated.

This introduces an additional set of Lagrangian multipliers. This is often referred to

as the dual variable approach, because each inequality has the potential of resulting
in two equalities: one for the upper limit and one for the lower limit.

However, the upper and lower limit cannot be simultaneously violated; thus, out of
the possible set of additional Lagrangian multipliers only one of the two will be
included at any given operating point and thus the dual limits are mutually
exclusive.
In this case, the additional equation is added:

P23 V2V3Y23 cos (θ 2 − θ3 − φ23 ) − V22Y=

23 cos φ23 0.100 per unit

This now increases the number of equality constraints to 3 (including the two original
powerflow equations)

Therefore an additional unknown state is required Choose PG3

 New Jacobian:

∂g1
 ∂g1

∂g1 ∂
=g 1

 ∂x1
(
V2 V1Y12 sin (θ 2 − θ1 − φ21 ) + V3Y13 sin (θ 2 − θ3 − φ23 ) )
 ∂x1 ∂x2 ∂x3 
 ∂g   ∂g 2 ∂g 2 ∂g 2  ∂g1
 ∂x  =  ∂x  =−V2V3Y32 sin (θ 2 − θ3 − φ23 )
   1 ∂x2 ∂x3  ∂x2
 ∂g3 ∂g3 ∂g3  ∂g
 
∂x3  ∂x = 0
1
 ∂x1 ∂x2
3
 ∂g 2
=−V3V2Y23 sin (θ3 − θ 2 − φ32 )
∂x1
∂g 2
= V3V1Y13 sin (θ3 − θ1 − φ31 ) + V2Y23 sin (θ3 − θ 2 − φ32 )
∂x2
∂g 2
=1
∂x3
∂g3
=−V2V3Y23 sin (θ 2 − θ3 − φ23 )
∂x1
∂g3
= V2V3Y23 sin (θ 2 − θ3 − φ23 )
∂x2
∂g3
=0
∂x3
 1 
 ∂g     ∂f 
 ∂u = 0  ;  ∂u = [1 + 0.025 PG 2 ]
   
0 

 ∂f   ∂C   ∂PG1   ∂C   ∂PG 3 
 ∂x     +  
∂
 G1  
P ∂ x ∂
  G3  
P ∂x 
V1V2Y12 sin (θ1 − θ 2 − φ12 ) 
 
= (1 + 0.125 PG1 ) V1V3Y13 sin (θ1 − θ3 − φ13 ) 
 0 

 V3V2Y32 sin (θ3 − θ 2 − φ32 ) 

 
+ (1 + 0.050 PG 3 ) −V3 (V1Y13 sin (θ3 − θ1 − φ31 ) + V2Y23 sin (θ3 − θ 2 − φ32 ) ) 
 
 0 
Proceeding with the steepest gradient algorithm iterations yields the final
constrained optimal generation scheduling:

 P1  141.3 
 P  =  490.3 MW with cost: $7,785/MWhr
 2  
 P3  321.2 

and P23 = 100.0 MW.

Only slightly different from penalty function approach
State Estimation

The state estimator is designed to give the best estimates of the voltages and phase
angles minimizing the effects of the measurement errors.

A set of specified measurements of a power system is said to be observable if the

entire state vector of bus voltage magnitude and phase angles can be estimated
from the set of available measurements.

An unobservable system is one in which the set of measurements do not span the
entire state space, and therefore not all states can be estimated.
Estimate the power system states and
use the chi-square test of inequality
with α = 0.01 to check for the presence
of bad data in the measurements.
 m

∑ w  z − h ( x )
2
Recall: Minimize e 2
= e ⋅e =
T
i i i
i =1

 x1  θ 2   1 
What is x?  x  = θ   0.010 
 2  3  
 x3  V3   1 
 0.050 
 
1
What is W? W = R −1 =  
 0.075 
 1 
 
 0.050 
 1 
 
 0.075 
What is h(x)?

V3 : h1 ( x) = x3
P13=
: h2 ( x) (
V1 x3Y13 cos ( − x2 − φ13 ) − V12Y13 cos φ13 )
Q21=
: h3 ( x) (V V Y
2 1 21 sin ( x1 − φ21 ) + V2 Y21 sin φ21
2
)
=
P3 : h4 ( x) ( 3 1 31 ( 2 31 ) 3 2 32 ( 2 1 32 ) 3 Y33 cos φ33
x V Y cos x − φ + x V Y cos x − x − φ + x 2
)
Q2=
: h5 ( x) (V2V1Y21 sin ( x1 − φ21 ) − V22Y22 sin φ22 + V2 x3Y23 sin ( x1 − x2 − φ23 ) )
 ) H xT R −1 [ z − h( x=
Set up equations: F ( x= )] 0
What is Hx ?

 0 0 1 
 
 0 V x Y
1 3 13 sin ( − x 2 − φ 13 ) V Y
1 13 cos ( − x2 − φ13 ) 
 V1V2Y21 cos ( x1 − φ21 ) 0 0 
 
 x3V2Y32 sin ( x2 − x1 − φ32 ) − x3V1Y31 sin ( x2 − φ31 ) V1Y31 cos ( x2 − φ31 ) 
 − x3V2Y32 sin ( x2 −x1−φ32 ) +V2Y32 cos ( x2 −x1−φ32 ) 
 
 +2 x3Y33 cos φ33 
 V V Y cos ( x − φ ) −V x Y cos ( x − x −φ ) V Y sin ( x − x − φ ) 
 1 2 21 1 21 2 3 23 1 2 23 2 23 1 2 23

 +V2 x3Y23 cos ( x1−x2 −φ23 ) 
The Newton-Raphson iteration to solve for the set of states x that minimize
the weighted errors is:

 ( ) ( )
 H xT x k R −1 H x x k   x=
 
k −1
− x k

 H T
x x ( )
k
R −1

 z − h ( x k
) 

Iteration 1: Starting with a flat start:

 1.0000   0 0 1.0000 
 0.0202   0 −10.0990 −1.0099 
   
h( x 0 ) =  −0.0664  and H x0 =  −0.2257 0 0
   
 −0.0198  −9.9010 20.0000 1.9604 
 −0.1914   −1.2158 0.9901 −9.9010 
  228.2791 θ 21   −0.0119 
   1  −0.0624 
and F ( x ) = −593.9313 θ 3  =
0
   
 −75.0229  V31   0.9839 
 
too large → not converged
Repeat iterations θ 2   −0.0113
   −0.0633
Repeating until convergence yields: x = θ3  =
 
V3   0.9851
 
 0.9851   0.975
 0.6668   0.668
   
and h =  -0.0636  compare with z = -0.082 
   
-1.1817  -1.181 
-0.0784  -0.086 
 To check for the presence of bad data, the weighted sum of squares of the
measurement errors are compared to the chi-square distribution for k = 2 and
α = 0.01. k = m-n = 5-3

The weighted sum of squares is

5
1
∑ σ ( z − h ( x) )
2
=f 2 i i
i =1 i

(−0.0101) 2 (0.0012) 2 (−0.0184) 2 (0.0007) 2 (−0.0076) 2

= + + + +
0.010 0.050 0.075 0.050 0.075
= 0.0155

This value is less than χ2,0.01 = 9.21; therefore, the data set is good and does
not contain any spurious measurements.